Merge OpenAI Triton commit 86a2ac7
#2630
Conversation
…odegen bug (#4873)" (#4973) After investigation of the differences caused by triton-lang/triton#4774 in the internal tests, we concluded that they were introduced by change in the layouts selected for the reduce operations. Re-introducing that change, as it is functionally correct and should be beneficial for performance.
This commit adds initial support for scaled_dot with mxfp8 LHS and fp8 RHS. It supports both mfma32 and mfma16 intrinsic variants. Right now we are missing software emulation for `Float8E4M3FN` type, so this only enables for `Float8E5M2`.
…`interpreter.cc` (#4976) `#include <atomic>` is already used in other triton files, so I believe it's not a cardinally change. Changes come from triton-lang/triton#4045
Signed-off-by: Anatoly Myachev <[email protected]>
|
@whitneywhtsang I am ending this and stopping this activity for now as agreed with you offline :) |
|
@whitneywhtsang ready for review |
|
Is the pass rate degradation solely due to test_scaled_dot? Can we open an issue to fix that? |
Yes, simply because the number of parameter combinations has increased, before this PR this test also did not work on XPU. Will open. |
Looks like to the number of test cases are unchanged, but this PR marks the failures as skipped instead of xfailed, that's why pass rate is affected. |
Ah, it increased only for AMD, I see. |
This PR change the Triton base from 1d5fdfe to 86a2ac7 (Oct 28).
Pass rate: 99.83%->97.41%
Please do not squash and merge this PR.